Non-projective Dependency-based Pre-Reordering with Recurrent Neural Network for Machine Translation Perform sentence reordering as a preprocessing step for machine translation using dependency trees and recurrent neural networks. Paper: Antonio Valerio Miceli Barone and Giuseppe Attardi. Non-projective Dependency-based Pre-Reordering with Recurrent Neural Network for Machine Translation. Association for Computational Linguistics, Stroudsburg, PA, USA. 2015. http://www.di.unipi.it/~miceli/rnnrm-cameraready.pdf Note: this is very experimental research code and is not particularly easy to use or documented. Usage: To train: 0. Parse your source corpus in the CoNNL format. Prepare your symmetrized alignments in the Moses format 1. Run depHintFromAlignmentOnaPap.py to add reordering information to your parse tree 2. Edit reordererState.py to make sure to import the proper feature extractor (Base-RNN and Base-GRU models can use the same feature extractors, while the Fragment-RNN model needs its own extractor denoted by the "_path" suffix). 3. Run preprocessTrain_base.py or preprocessTrain_path.py (depending on which kind of model you are using) to generate a feature dictionary from the training corpus 4. Edit one of the trainer_X.py scripts to make sure it uses the proper model type and hyperparameters and run it To reorder: 0. Parse your source corpus in the CoNNL format. 1. Edit one of the decoder_X.py scripts to make sure it uses the same model type and hyperparameters as the trainer_X.py and run it. 2. Run unparse-by-hints.py to extract the words in the permuted order from the decoded parse tree